Creating artificial intelligence agents capable of researching, using tools, and completing long tasks no longer requires using the most powerful model at every step. OpenAI introduces GPT-5.6 with a concrete promise for startups and product teams: better performance, lower costs, and new ways to organize how agents work.
More capability with less spending
The GPT-5.6 family aims to solve one of the biggest challenges in AI applications: complex tasks may require many model calls, large amounts of context, and high token consumption.
OpenAI says GPT-5.6 maintains or improves the performance of previous generations while operating more efficiently. One of the key changes is that smaller models in the family, such as Luna and Terra, can achieve results similar to GPT-5.4 and GPT-5.5 on certain tasks when they are allowed to use more reasoning capacity.
What’s the practical consequence? A company can reserve its most advanced model for difficult decisions and use more affordable models to extract data, classify documents, or perform repetitive steps.
For example, a legal startup could use Terra or Luna to read handwritten memos and extract information before sending the data to an agent responsible for legal analysis. This way, the more expensive model focuses on interpreting and deciding instead of processing every line from the beginning.
An example of savings in complex searches
OpenAI compares the performance of GPT-5.5 and GPT-5.6 Luna on BrowseComp, a test that evaluates the ability to find uncommon information through searches.
GPT-5.5, with a very high reasoning level, achieved a score of 84.36% at a total cost of $33.27. GPT-5.6 Luna achieved practically the same result, 84.04%, for $1.33 at launch. The company says it later reduced prices even further.
This does not mean Luna will always replace larger models. It means the model choice can be made more precisely, depending on the type of task, the volume of requests, and the level of reasoning required.
Three tools for more efficient agents
In addition to improving its models, OpenAI added new functions to the Responses API. These functions aim to prevent an agent from repeating work, wasting context, or using the model for tasks that a program can handle more effectively.
1. Preserve reasoning between turns
Agents can retain part of the work completed from one interaction to the next. They can also use native compaction to summarize long conversations without losing continuity in the task.
This matters when an agent works for a long time. Instead of rebuilding the entire context at every step, it can continue from a previous state and maintain a more coherent line of work.
In an ARC-AGI-3 test, GPT-5.6 Sol went from a score of 13.3% to 38.3% when retained reasoning and compaction were activated. The result was achieved without changing the model and while using approximately six times fewer output tokens.
2. Divide tasks among multiple agents
Complex tasks do not always have to be solved sequentially. With multi-agent orchestration, a primary agent can distribute the work among several specialized agents.
One can research sources, another can compare results, and a third can review inconsistencies. Then, the primary agent brings the answers together and prepares a final summary.
This approach can reduce execution time and improve quality for work that can be divided in parallel. You can also tell the model when it should create subagents, which is useful for avoiding unnecessary expenses on simple tasks.
3. Leave mechanical work to code
Not everything that happens inside an agent requires reasoning. If the system needs to consult 100 documents, filter them by date, and add up transactions, doing everything within the model’s context can be slow and expensive.
The so-called Programmatic Tool Calling allows GPT-5.6 to write JavaScript to coordinate tools, execute calls in parallel, and process results outside the context window. The model can then focus on what actually requires intelligence: interpreting information and making decisions.
The central idea is simple: code moves, filters, and combines data; the model analyzes what requires judgment.
Prompt caching also gets better
GPT-5.6 extends the prompt cache lifetime to a minimum of 30 minutes across the entire model family. It also allows you to deterministically set cache points within the context.
This can increase the number of requests that reuse instructions and data that have already been processed. For applications with long prompts or repetitive workflows, a better reuse rate can reduce latency and cost.
OpenAI also recommends using prompt_cache_key correctly, a key that helps direct requests with the same prefix to the appropriate inference engine.
Architecture matters as much as the model
The main lesson from this update is that building with AI is not only about choosing the model with the highest score. The way you divide a task, preserve context, use multiple agents, and move operations to code can radically change the final result.
For a startup, this opens up an interesting possibility: creating more capable products without paying the price of using a frontier model at every step. The challenge is no longer just finding the most intelligent model, but designing a system that uses each resource where it actually adds value.
GPT-5.6 points precisely in that direction. Artificial intelligence becomes more useful when the model reasons better, but also when it learns not to spend reasoning on something that a rule, a function, or an automated process can handle.